diff --git a/lmms_eval/api/samplers.py b/lmms_eval/api/samplers.py
index 6683bba96..e1a3dda21 100644
--- a/lmms_eval/api/samplers.py
+++ b/lmms_eval/api/samplers.py
@@ -118,7 +118,7 @@ def add_question(self, doc, data_frame=None, index=None):
         if visual:
             self.contexts.append(visual)
         self.contexts.append(QAPairs(question))
-        self.contexts.append(self.target_delimiter)
+        # self.contexts.append(self.target_delimiter)
 
     def get_text(self, *, image_tokens="<image>", lazy=True):
         texts = []
diff --git a/lmms_eval/api/task.py b/lmms_eval/api/task.py
index cc3599622..0631fed9d 100644
--- a/lmms_eval/api/task.py
+++ b/lmms_eval/api/task.py
@@ -791,7 +791,7 @@ def fewshot_context(self, doc_id, num_fewshot, split):
         labeled_examples = Context(self, self.config.fewshot_delimiter, self.config.target_delimiter, self.config.description)
         if num_fewshot != 0:
             labeled_examples.extend(self.sampler.get_context(doc, num_fewshot))
-        labeled_examples.add_question(doc)
+        labeled_examples.add_question(doc, self.test_docs(), doc_id)
         return labeled_examples
 
     def apply_filters(self):
diff --git a/lmms_eval/models/llava.py b/lmms_eval/models/llava.py
index ea5e7d03e..a830fd8d9 100644
--- a/lmms_eval/models/llava.py
+++ b/lmms_eval/models/llava.py
@@ -349,6 +349,7 @@ def _collate(x):
                         answer = obj.answer
                         conv.append_message(conv.roles[0], question)
                         conv.append_message(conv.roles[1], answer)
+                        num_image_tokens = 0
 
                 # conv.append_message(conv.roles[0], question)
                 # conv.append_message(conv.roles[1], None)
@@ -358,14 +359,17 @@ def _collate(x):
             # The above for loop has bugs. When there is no visuals, e.g. pure text,
             # there will be no for loop execute resulting in an empty question_input (because no visuals)
             # Scenario 1 won't even be execute
-            if len(flattened_visuals) == 0:
-                for context in contexts:
-                    question = context
-                    conv = conv_templates[self.conv_template].copy()
-                    conv.append_message(conv.roles[0], question)
-                    conv.append_message(conv.roles[1], None)
-                    prompt_question = conv.get_prompt()
-                    question_input.append(prompt_question)
+            # if len(flattened_visuals) == 0:
+            #     for context in contexts:
+            #         question = context
+            #         conv = conv_templates[self.conv_template].copy()
+            #         conv.append_message(conv.roles[0], question)
+            #         conv.append_message(conv.roles[1], None)
+            #         try:
+            #             prompt_question = conv.get_prompt()
+            #         except Exception as e:
+            #             pass
+            #         question_input.append(prompt_question)
 
             # input_ids = tokenizer_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to(self.device)
             # preconfigure gen_kwargs with defaults